Research map / Graph learning / Graph self-supervised learning
MoE-Guided Graph Diffusion for Oriented Molecule Design
Shuochen Li, Xiangqi Guo, Huobin Tan, Lei Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RL finetuning after MoE-guided pretraining (ablation) | MEGD · QM9 full-sample | Gap: 7.107 -> 9.110 | Improved |
| add RL finetuning (ablation) | MEGD · ZINC250k full-sample | Gap: 5.465 -> 7.859 | Improved |
| add timestep input to the MoE gating network (ablation) | MEGD · three target properties after RL finetuning | performance: 17.8% drop when removed | Improved |
| add sub-structure input to the MoE gating network (ablation) | MEGD · three target properties after RL finetuning | performance: 26.1% drop when removed | Improved |
| use both timestep and sub-structure gating inputs (ablation) | MEGD · three target properties after RL finetuning | performance: 41.4% drop when both removed | Improved |
| replace GDM layers with MoE layers (pretraining only) (ablation) | graph diffusion model · QMOF | validity (%): 63.669 -> 72.315 | Improved |
| replace GDM layers with MoE layers (pretraining only) (ablation) | graph diffusion model · QM9 | validity (%): 99.498 -> 99.220 | No clear effect |
| select two experts instead of one (ablation) | MEGD · hyperparameter tuning, 6 total experts, balance weight 0.01 | target property score: 9.069 -> 9.926 | Improved |
| increase total experts from 6 to 8 (two selected) (ablation) | MEGD · hyperparameter tuning | target property score: 9.926 -> 9.080 at balance weight 0.01 | Hurt |
| use a large load-balance weight (1) instead of moderate (ablation) | MEGD · hyperparameter tuning, 2 of 8 experts | target property score: 9.080 -> 8.357 | Hurt |
| MEGD vs best RL-guided diffusion baseline GDPO | MEGD · QM9, ZINC250k, QMOF full-sample | Gap: 8.110 -> 9.110 (QM9) | Improved |
| MEGD vs MOOD on top-25% ZINC250k samples | MEGD · ZINC250k top-sample | Gap: 11.361 vs 11.296 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiGress — DiGress: Discrete Denoising diffusion for graph generation (ICLR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design (NeurIPS 2025 · similarity 0.88)
- MACS: Multi-Agent Reinforcement Learning for Optimization of Crystal Structures (NeurIPS 2025 · similarity 0.87)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.87)
- MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models (NeurIPS 2026 · similarity 0.87)
- Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.87)
- MANDREL: Modular Reinforcement Learning Pipelines for Material Discovery (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
